Geospatial Artificial Intelligence (GeoAI) — the convergence of satellite-derived data with modern machine-learning techniques—has become a decisive factor in global disaster-risk governance. In little more than five years, the active Earth-observation satellite fleet has tripled, and open-weight vision models now convert petabytes of imagery into decision-ready layers in minutes. For the United Nations Platform for Space-based Information for Disaster Management and Emergency Response (UN-SPIDER) and its network of twenty-seven Regional Support Offices
(RSOs), this technological acceleration marks a transformation, not merely an upgrade: hazards can now be detected in near real time, their impacts forecast weeks in advance, and policy responses grounded in defensible evidence rather than retrospective estimates.
This Compendium presents the first global stock-take of GeoAI practice across the UN-SPIDER community. Between March and May 2025, every RSO completed a structured submission template describing its most advanced workflows. After peer review and validation, eighteen best-practice cases were selected that together span rapid-onset hazards such as floods, wildfires and earthquakes, as well as slow-onset stresses including drought, coastal erosion and urban heat. Across these cases, the median time required to generate an operational product fell by 65 to 85 per cent relative to pre-GeoAI methods, while pixel-level accuracy improved by 8 to 15 percentage points. Such gains are not abstract: they directly advance Sendai Framework targets on mortality and economic loss, strengthen national adaptation planning under Article 7 of the Paris Agreement, and improve reporting on Sustainable Development Goal indicators 11.5, 13.1 and 9.5.
The Compendium’s evidence reveals both momentum and persistent gaps. On the positive side, all RSOs now operate at least one containerised deep-learning pipeline on commercial or sovereign clouds; three RSOs have already integrated GeoAI outputs into national early-warning chains, trimming evacuation lead times by up to six hours in cyclone-prone deltas. Open-weight foundation models such as Prithvi and Satlas—coupled with disaster response cloud credits—have lowered entry barriers to the point where analysts in low-income states can match the analytic throughput of space-faring nations. Yet data inequity, compute scarcity, model drift, and unresolved ethical issues remain significant obstacles, especially for small-island and land-locked developing States where ground truth is sparse and data-sovereignty laws tight.
To close these gaps, the report recommends an integrated programme of action for the period 2025-2028. At the centre of that programme is a GeoAI Partnership Portal hosted on the UN-SPIDER Knowledge Platform: a one-stop registry of open models, pooled cloud-compute quotas and focal-point contacts that will streamline collaboration among RSOs, governments, academic laboratories and private-sector providers. Complementing the portal, UN-SPIDER will curate a library of containerised reference pipelines—flood mapping, landslide detection, drought monitoring—complete with standardised model cards and licensing metadata so that any agency can deploy proven workflows with minimal configuration. Localisation is non-negotiable: every deployment will reserve a portion of indigenous ground truth for validation, and RSOs will organise community labelling campaigns and translate critical documentation into local languages to secure contextual fit.
Looking to the horizon framed by COP 30 (Belém, November 2025), the 2030 Sendai stock-take and the SDG Summit of 2027, GeoAI is poised to evolve in three fundamental directions. First, ensemble learning pipelines will move the field from historical damage mapping to sub-seasonal impact forecasts, giving governments the foresight to pre-position relief goods, adjust cropping calendars and plan climate-resilient infrastructure. Second, real-time analytics will feed directly into national policy dashboards, serving climate-budget tagging and loss-and-damage accounting under Article 8 of the Paris Agreement. Third, privacy-preserving federated learning will allow models to improve across borders without raw data ever leaving sovereign clouds—a critical capability for states with strict residency laws.
The Compendium’s overall conclusion is unequivocal: GeoAI now sits at the heart of UN-SPIDER’s mandate to make space-based information accessible, actionable and equitable. The distributed RSO model—sovereign data held locally, shared algorithms circulated globally, federated validation applied regionally—has proven its worth in translating cutting edge science into life-saving practice. Institutionalising the actions set out in this report will ensure that GeoAI becomes not a boutique research interest but a durable, ethically governed backbone for disaster resilience and sustainable development at precisely the moment the planet’s risk landscape demands it most.
